For decades, the primary challenge of robotics has not been the ability to move, but the ability to feel. While modern robots can map a room in milliseconds using LiDAR or identify a thousand different objects via computer vision, they remain largely numb. This sensory gap creates a critical failure point in what engineers call dexterous manipulation
—the ability to handle fragile objects, operate tools, or navigate tight spaces without causing damage.
Hong Kong-based DAIMON Robotics is attempting to bridge this gap by treating touch not as a secondary supplement to vision, but as a primary modality of intelligence. The company recently released Daimon-Infinity, which it describes as the largest omni-modal robotic dataset for physical AI. By combining high-resolution tactile sensing with a massive array of real-world tasks, the initiative aims to provide the necessary fuel
to move embodied AI from controlled laboratory demos to unpredictable human environments.
The project is a global collaborative effort, involving partners such as Google DeepMind, Northwestern University and the National University of Singapore. By open-sourcing 10,000 hours of this data, DAIMON is attempting to address the industry-wide bottleneck of data scarcity, specifically the lack of physical interaction data required for robots to operate safely and effectively in the real world.
This shift toward robotic tactile sensing represents a fundamental change in how AI “brains” are trained. For most of the current generation of humanoid robots, the dominant architecture is the Vision-Language-Action (VLA) model. Yet, DAIMON is pioneering a fresh framework: Vision-Tactile-Language-Action, or VTLA. This architecture elevates tactile feedback to a status on par with vision, allowing a robot to sense deformation, slip, and friction in real time.
The Engineering of Touch: Monochromatic Vision-Based Sensing
At the heart of DAIMON’s approach is a specialized piece of hardware: a monochromatic, vision-based tactile sensor. Unlike traditional sensors that might rely on simple pressure plates, this technology packs over 110,000 effective sensing units into a module the size of a human fingertip. The sensor works by capturing a sequence of visual images of the deformation on the fingertip’s surface when it touches an object.
By analyzing these images, the system can infer precise forces, material properties, and surface textures. This vision-based approach is strategic; since the tactile data is captured as an image, it integrates seamlessly into the existing visual frameworks used by VLA models, facilitating the transition to the VTLA architecture.
According to Prof. Michael Yu Wang, co-founder and chief scientist at DAIMON Robotics, this technical path was chosen to mimic the physiological capabilities of human fingertips. The goal is to enable robots to perform tasks that are nearly impossible without touch, such as locating objects in total darkness or detecting the exact moment a glass object begins to slip from a grip.
Daimon-Infinity and the Data Bottleneck
The release of the Daimon-Infinity dataset is a response to a persistent problem in embodied AI: the scarcity of high-quality interaction data. While large language models (LLMs) can be trained on the vast expanse of the internet, robots require physical data—actual records of how a finger interacts with a surface—to learn. DAIMON has addressed this by building a distributed, out-of-lab data collection network capable of generating millions of hours of data annually.
The resulting dataset spans a wide range of scenarios, from domestic chores like folding laundry to complex industrial manufacturing on assembly lines. By partnering with institutions like the National University of Singapore and enterprises like China Mobile, DAIMON has gathered practical, application-driven data from over 80 real scenarios and more than 2,000 human skills.
“To drive the advancement of the entire embodied AI field, we have open-sourced 10,000 hours of the dataset for the broader community.” Prof. Michael Yu Wang, DAIMON Robotics
From Academia to the “3D” Business Model
The strategic direction of DAIMON is heavily influenced by the academic background of Prof. Michael Yu Wang. With roughly four decades of experience in the field, including a PhD from Carnegie Mellon University and the founding of the Robotics Institute at the Hong Kong University of Science and Technology, Wang has spent years studying the nuances of manipulation.
Transitioning from academia to a startup, Wang and co-founder Dr. Duan Jianghua have implemented what they describe as a 3D
business model: Devices, Data, and Deployment. This vertical integration ensures that the hardware (Devices) is optimized to collect the necessary information (Data), which is then used to train models for specific real-world environments (Deployment).
This model is designed to move beyond the impressive demo
phase that characterizes many robotics companies. Instead of showing a robot performing a single task in a lab, DAIMON is focusing on “embodied skills”—the ability for a robot to operate autonomously in unstructured environments, make decisions based on touch, and act as a reliable partner for humans.
The Path to Real-World Deployment
While the road to general-purpose humanoid robots remains long, DAIMON identifies specific sectors where touch-enabled robots are most likely to achieve large-scale deployment first. One primary example is the hospitality industry in China, where delivery robots are already nearly 100 percent deployed in major hotels for simple navigation tasks.
The next frontier involves more complex environments, such as overnight drugstores and convenience stores. In these settings, robots must navigate densely packed shelves and reach into tight spaces to retrieve objects. Current parallel grippers often fail in these scenarios; however, a robot equipped with three slim, tactile-enabled fingers can touch and roll an object toward itself, mimicking human behavior.
the convergence of high-torque electric hardware and VTLA-driven intelligence is expected to produce robots that can safely navigate human homes. By integrating tactile feedback, these machines can move from being simple tools to becoming autonomous assistants capable of handling the fragility and unpredictability of daily life.
Key Technical Comparisons: VLA vs. VTLA
| Feature | Vision-Language-Action (VLA) | Vision-Tactile-Language-Action (VTLA) |
|---|---|---|
| Primary Inputs | Visual images, Natural language | Visual images, Tactile data, Natural language |
| Feedback Loop | Relies on visual confirmation | Real-time physical contact feedback |
| Handling Fragility | High risk of damage/slip | Precise force control and slip detection |
| Environment | Requires clear line-of-sight | Capable of operation in dark/occluded spaces |
As DAIMON Robotics continues to expand its distributed data collection network and refine its monochromatic sensors, the industry will be watching for the first large-scale deployment of humanoid robots in retail environments. Further updates on the Daimon-Infinity dataset and its impact on open-source physical AI are expected as more researchers integrate the 10,000 hours of open data into their own models.
What do you think about the prospect of “feeling” robots in your home or local store? Share your thoughts in the comments below or join the conversation on our social channels.